An intelligent surface segmentation approach based on U-Net for structured and freeform surface characterisation

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review


With the freeform structured surface becoming more dominant in the engineering surface metrology and manufacturing as they have deterministic patterns which designed with some specific functionalities to meet the engineering requirements such as optical, electric contact and bearing properties, the ability to adequately characterise them is crucial for optimising the performance through reducing cost and achieving precise control of such specifically functional components. A general surface characterisation scheme for complex freeform includes three operations: form removal, denoising, segmentation. Following the first two steps, the surface measurement will be converted to a scale-limited surface. The next crucial step is segmentation, which separates the surface topography into a number of non-intersecting regions so that they can be analysed separately and in relation to one another, for example, by computing shape attributes or its pertinent dimensions. Traditional computer-vision methods such as watershed and active contour approaches perform well on this task, but these algorithms have high computational complexity with long running time and require test phase for fine-tuning and rely on users-judgement. Additionally, inappropriate initial conditions will lead to over- or under- segmentation. Deep learning-based segmentation techniques become dominated with outstanding performance and higher accuracy. Therefore, we proposed a novel deep learning-based surface segmentation method for the freeform structured surfaces which based on the U-Net model and data augmentation techniques are utilised to enlarge the raw dataset of twenty surface measurements. The training data include converted RGB-images of surfaces and corresponding feature masks which is the ground truth pixel labels generated using computer vision techniques including thresholding and edge operators. Once the model has been trained, it can output the drawn feature map of input surface with precise boundaries efficiently. The U-Net segmentation model is a kind of Encoder-decoder architecture with benefits of lower cost and high efficiency with great segmentation accuracy using few training data. The experimental results show the remarkable and applicable performance of structured surface segmentation to meet the metrological requirements, which can support the intelligent surface characterisation framework and for further feature attributes analysis and parameterisation.
Original languageEnglish
Title of host publication23rd International Conference & Exhibition for European Society for Precision Engineering and Nanotechnology
EditorsO. Riemer, C. Nisbet, D. Phillips
Number of pages4
ISBN (Print)9781998999132
Publication statusPublished - 12 Jun 2023
Event23rd International Conference & Exhibition for European Society for Precision Engineering and Nanotechnology - Technical University of Denmark, Copenhagen, Denmark
Duration: 12 Jun 202316 Jun 2023
Conference number: 23


Conference23rd International Conference & Exhibition for European Society for Precision Engineering and Nanotechnology
Abbreviated titleEUSPEN 2023
Internet address

Cite this